Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer built jevchat, an experimental chatbot that turns the Jev API into a text generator by repeatedly asking it which symbol comes next and sampling from the returned probability distribution. The tool, described…
A September 2026 survey of self-hosted inference orchestrators compared LocalAI, exo, GPUStack, vLLM, Ollama, llama.cpp, LiteLLM, Xinference, NVIDIA Dynamo/llm-d, SkyPilot and dstack across modalities, multi-machine supp…
Ds4 now supports directional steering, a runtime activation edit that applies a normalized f32 direction per transformer layer during inference, with steering files shaped 43x4096 for DeepSeek V4 Flash, 45x4096 for GLM 5…
A developer has published a CI-runnable script that fetches a gateway's pricing.usd.json file and flags near-duplicate model identifiers whose cache-read prices diverge, after finding that strings like claude-fable-5 and…
Northwestern University researchers have demonstrated that the standard method for detecting training-data contamination in AI forecasting benchmarks is mathematically flawed, after a test they built falsely accused four…
A developer trained a roughly 14-million-parameter small language model on the TinyStories dataset and paired it with Jev, a model that judges text by answering yes/no questions with calibrated confidence, to improve its…
An AI model called GPT-6 Astra has reportedly deciphered a previously unbroken World War I German radio message encoded with the ADFGVX cipher, according to PrinzAI. The message, sent on November 27, 1918, used the codew…
A developer identified and fixed a KV-cache offloading bug in vLLM that caused incorrect chunking for models with mixed KV-cache groups. The existing implementation assumed a single KV-cache group layout and relied solel…
In a 2019 Slate Star Codex post, Scott Alexander argued that OpenAI's GPT-2 is not merely a "brute-force statistical pattern matcher" but performs the same kind of concept-level pattern recognition as humans, differing o…
Huawei unveiled the Atlas 960E SuperPoD at HUAWEI CONNECT 2026 in Shanghai on September 17, an AI cluster that links 4,096 Ascend neural processing units into a single logical machine with unified memory addressing, targ…
Microsoft AI chief Mustafa Suleyman said on CNN's Fareed Zakaria GPS that concerns about China's AI progress should not be used as an argument against regulating the technology, saying "I don't think we should use China …
LangChain tested TypeSafe AI's Jev, a typed classifier rather than a text-generating LLM, as an agent-eval judge, and Jev matched a human reviewer on all 500 repeated decisions while costing a fraction of a cent per call…
TypeSafe AI launched Jev on September 15, a "System One" model that generates zero text and returns typed answers with probabilities in 70 to 500 milliseconds, priced at $0.042 per million input tokens with output free. …
Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, two native speech-to-speech models that process speech directly and execute tools asynchronously while a conversation continu…
Sebastian Raschka argued in a blog post that Jev, the system-one model from Typesafe.ai, should not be dismissed as "just a classifier," crediting its generalization across tasks such as classifying emails, playing video…
Pirate Face launched a decentralized peer-to-peer mirroring service that converts open models from Hugging Face — including LLMs, image, audio, and datasets — into checksum-verified torrents, with each file carrying its …
A developer building an autonomous coding system with Claude Code split work between a planner agent and fresh-context implementer sub-agents, but found that roughly one in three implementer runs solved the wrong problem…
The Federal Register's document search interface briefly listed two "Qwen3:0.6B" search modes — a large language model from Alibaba Cloud — before both options were removed roughly a day after a Switzerland-based commodi…
Ternary Bonsai 2 27B, a ternary-weight multimodal model built on a Qwen3.8 27B base, ships at roughly 6GB with weights restricted to {-1, 0, +1} at 1.76 bits per weight with FP16 group scales, according to the model's pr…
A developer released pi-jev-router, a minimal Pareto-optimal OpenRouter model router for the pi coding agent that classifies tasks with a single Jev call and assigns models per role: anthropic/claude-sonnet-5 for plannin…